2016-09-07

## Concurrency models in CSP

Preliminaries

We’ll start by looking at the weakest concurrency model covered in the paper, Read Atomic. In fact, to keep things really simple to start with, we’re only going to look at one of Read Atomic’s two axioms: internal consistency. (We’ll look at external consistency in the next post.)

A transaction is internally consistent if it “reads its own writes”. This is the simplest axiom covered in the paper, since it expresses a property that’s strictly local to each transaction; we don’t have to consider the behavior of any other transaction when defining what it means for a transaction to be internally consistent.

#### Types and channels

Our goal is to construct a CSP process that describes internal consistency. We start by defining some data types and channels.

## Channels

Each $$\textbf{channel}$$ statement defines one kind of event that your processes can use to communicate. Each event will typically carry some data; you use $$\textbf{datatype}$$ and $$\textbf{nametype}$$ statements to define the different kinds of data that can be carried by those events.

We’ll first need to talk about the objects in the data store, and the values that those objects can have:

$\textbf{nametype} ~ \textsf{Object} = \{1..2\} \\ \textbf{nametype} ~ \textsf{Value} = \{1..5\}$

An $$\textsf{Object}$$ is the name of one of the objects that can be read or written. The $$\textsf{Value}$$ type is the set of values that can be stored in each object. We keep both of these sets small, to make it easier for FDR to perform refinement checks without needing too much memory. (In particular, we can’t use $$\mathbb{Z}$$, the infinite set of integers, like they do in the original paper, since that definitely won’t fit into memory!)

Once we have our data types defined, we can define our channels:

$\textbf{channel} ~ \texttt{read} : \textsf{Object} \times \textsf{Value} \\ \textbf{channel} ~ \texttt{write} : \textsf{Object} \times \textsf{Value}$

A $$\texttt{read}$$ event tells us that a particular value was read from a particular object, while a $$\texttt{write}$$ event tells us that a particular value was written to a particular object. For instance, $$\texttt{read}{.}1{.}3$$ tells us that a transaction read the value 3 from object 1, and $$\texttt{write}{.}2{.}4$$ tells us that a transaction wrote the value 4 to object 2.

## Timing

Note that we don’t have to include timestamps or anything like that in our channel descriptions to indicate that one event occurred before another. In CSP, we get that for free from the operators that we use to construct a process. For instance, the “prefix” operator ($$\rightarrow$$) lets us say something like:

$\texttt{read}{.}1{.}3 \rightarrow \texttt{write}{.}2{.}4 \rightarrow \ldots$

meaning that a transaction read the value 3 from the object 1, and then wrote the value 4 to the object 2 (and then does something else that we don’t care about right now).

#### Undefined objects

Now we can finally start defining some processes! Within a transaction, each object starts undefined, meaning that this transaction has not yet written a value to the object. (Remember that internal consistency does not take into account what values other transactions write into an object.) We can construct a process (parameterized by an $$\textit{obj}$$ variable) that remembers that one particular object is currently undefined:

$\textrm{Undefined}(\textit{obj}) = {} \\ \quad\texttt{write}\,!\textit{obj}\,?\textit{value} \rightarrow \textrm{Defined}(\textit{obj}, \textit{value}) \\ \qquad \Box \\ \quad\texttt{read}\,!\textit{obj}\,?\textit{value} \rightarrow \textrm{Undefined}(\textit{obj})$

The first clause tells us that you can write to an undefined object. The $$!\textit{obj}$$ part ensures that we only worry about $$\texttt{write}$$ events for this object; the $$?\textit{value}$$ part means that we don’t want to place any restrictions on what value you can write into the object. If you write into an undefined object, the object becomes defined (which we’ll consider below), and we remember the value that was written.

## $$\rightarrow$$ (prefix), $$?$$ and $$!$$

The prefix operator lets us introduce sequencing; $$a \rightarrow B$$ means that the process performs an event $$a$$, and then it does whatever is described by the process $$B$$. Nice and simple! The only wrinkle is that the left side of the arrow is an event, while the right side is a process.

You will usually see the special $$?$$ and $$!$$ event constructors used with the prefix operator. This is just some syntactic sugar that means we don’t have to explicitly write out separate clauses for $$\texttt{read}{.}1{.}1$$, $$\texttt{read}{.}1{.}2$$, $$\texttt{read}{.}2{.}1$$, $$\texttt{read}{.}2{.}2$$, and so on.

The $$?$$ constructor lets you create a new variable which can take on any of the possible values that could appear in that “slot” of the event; you can refer to the new variable in the rest of the clause if you need to know which particular value was selected.

The $$!$$ constructor lets you assert that only one specific value is allowed; that value is given by the variable that follows the $$!$$. (That means that $$?$$ should be followed by a new variable name, but $$!$$ should be followed by an existing variable name.)

The second clause tells us that you can also read from an undefined value. This might be surprising! Internal consistency doesn’t tell you which value you’ll read, so why should we include this case? This is one of the first hurdles you’ll encounter when learning about formal methods: in the world of math, there’s a very big difference between disallowing something and not caring about it. Here, we don’t care: just because a value hasn’t been defined by this transaction doesn’t mean that some previous transaction hasn’t committed a perfectly good value for it; in fact, all of the later concurrency models will be interesting exactly because of all of the different ways they decide which value you see! We need to include the $$\texttt{read}$$ clause, and use the $$?\textit{value}$$ syntax to not place any constraints on what value is read, to leave open all of those different interesting possibilities.

## The environment

When describing the events in a process, we always keep in mind that the process is going to communicate or interact with something else, called its environment. CSP uses a “rendezvous model”, which means that both sides of an interaction must agree before an event can take place. Typically, that “something else” is another process that we’ve combined together using one of CSP’s parallel composition operators. But we can also talk about an “observer” — such as you, the specification author — who wants to explore the behavior of the process by walking through all of its possible combinations of events.

## Alphabets

Each process has an alphabet, which is the set of events that it cares about. If an event is not in a process’s alphabet, the process cannot perform that event, and it cannot prevent any other processes from performing the event. In fact, the process isn’t even allowed to mention the event.

On the other hand, if an event is in a process’s alphabet, then the process controls whether and when that event occurs — or doesn’t! If the environment wants to perform an event in the process’s alphabet, but the process isn’t ready to perform that event, then the environment is prohibited from performing the event.

With great power comes great responsibility! You must be careful not to overspecify your processes; it can be easy to accidentally prevent an event from occurring, when you actually want to express that you don’t care whether it occurs. Ideally you do that by not including the event in your alphabet. But you might need the event in the your alphabet so that you can control it at some other time; you just don’t want to right now. In that case, you have to include a “fallthrough” clause like we did for reading undefined objects.

The two clauses are combined using external choice, which means that whenever it’s time for the transaction for perform another operation, it’s free to perform either a read or a write.

## $$\Box$$ (external choice)

CSP’s “choice” operators combine two processes, allowing one or the other to proceed, but not both. External choice ($$\Box$$) gives the environment full control over which “branch” occurs. For $$A \mathrel{\Box} B$$, we’re willing to act like process $$A$$ or process $$B$$, and don’t care (and have no way to influence) which one occurs; the environment decides.

#### Defined objects

Once we’ve written a value to an object (in this transaction), the object becomes defined. We use a parameterized process for defined objects, just like we did for undefined objects, but we need an extra $$\textit{currentValue}$$ parameter to keep track of the object’s current value.

$\textrm{Defined}(\textit{obj}, \textit{currentValue}) = {} \\ \quad\texttt{write}\,!\textit{obj}\,?\textit{newValue} \rightarrow \textrm{Defined}(\textit{obj}, \textit{newValue}) \\ \qquad \Box \\ \quad\texttt{read}\,!\textit{obj}\,!\textit{currentValue} \rightarrow \textrm{Defined}(\textit{obj}, \textit{currentValue})$

The first clause tells us that you can write to a defined object (just like you can write to an undefined one). Doing so should overwrite the current value with the new value; we handle this by using the new value as the parameter in our recursive call.

The second clause tells us that you can read from a defined object. Unlike the undefined case, internal consistency constrains what value you read when you do so. The $$!\textit{currentValue}$$ part enforces that the value returned is the current value of the object, not allowing any other $$\texttt{read}{.}\textit{obj}$$ events to occur at this point. Because of CSP’s rendezvous model, even if another process tries to read some other value, the fact that we’ve actively prevented it means that they won’t be able to.

#### Combining it together

We now have subprocesses telling us how internal consistency behaves, on a per-object basis, for both undefined and defined objects. The last thing we need to do is combine all of the per-object subprocesses together into a single process describing internal consistency as a whole.

## Parallel composition

The most important operators in CSP are the parallel composition operators, which combine two or more processes together and let them proceed concurrently. There are several different flavors of parallel composition, which we’ll see as we work through the concurrency models, but they all have the same basic “shape”.

Each of the processes that you’re combining are allowed to proceed concurrently. There is a set of events (called the “interface”) that the processes can use to communicate with each other. The parallel composition operators are different only in how they decide which events are included in the interface.

For events outside the interface, the processes are completely independent: they can proceed as quickly or as slowly as they want, and cannot interfere which each other in any way. If two processes both want to perform the same non-interface event, they can. However, those events are also completely independent: the environment will see both “instances” of the event occur, and can’t tell which one belonged to which process.

Events in the interface, however, are synchronous: all of the processes being combined must be ready to perform the event before any of them are allowed to. Once all processes are ready, the event can occur. If the event occurs, it’s shared; the environment will only see one “instance” of the event, and all of the processes will simultaneously “handle” the event and move on to the their next state.

$\textrm{InternalConsistency} = \textstyle\Interleave_{obj} \textrm{Undefined}(obj)$

The $$\Interleave_{obj}$$ notation is kind of like the $$\sum$$ operator that you might know from plain old arithmetic; it means that we’re going to instantiate the $$\textrm{Undefined}$$ subprocess once for each possible $$\textit{obj}$$, and then combine them all together using the “interleave” operator ($$\interleave$$). Each object starts undefined, so we only have to mention the $$\textrm{Undefined}$$ subprocess; each subprocess will automatically transition to $$\textrm{Defined}$$ at the right time, if the corresponding object is ever written to.

## $$\interleave$$ (interleaving)

The simplest parallel composition operator is “interleaving”, where the interface set is empty. All of the processes execute completely independently, and cannot communicate with each other, even if their alphabets have any events in common.

We can use interleave here because the alphabets of our subprocesses are disjoint — that is, they all work with completely distinct sets of events. (This is what we want, since whether or not object 1 is defined should have no bearing on what we can read or write for object 2.) No matter how we try to combine these subprocesses together, there is no way that they can interact with each other, so we might as well use the simplest composition operator.

#### Testing it out

We now (presumably!) have a working CSP process describing the internal consistency axiom from the paper. If my running commentary isn’t enough to convince you that the process is correct, we can also use FDR to perform some tests! This is very preliminary; once we have the full consistency model specified, along with the reference implementation from the paper, we’ll do some more thorough (and interesting!) checks.

Our goal is to have FDR check whether the following transactions are valid:

$\textrm{UnrepeatableRead} = \texttt{write}{.}1{.}1 \rightarrow \texttt{read}{.}1{.}2 \rightarrow \textrm{STOP}$

$\textrm{RepeatableRead} = \texttt{write}{.}1{.}1 \rightarrow \texttt{read}{.}1{.}1 \rightarrow \textrm{STOP}$

The first CSP process represents an “unrepeatable read”: the transaction writes the value 1 into object 1, but when it then tries to read from object 1, it somehow gets the value 2. This should be outlawed by internal consistency. The second CSP process should be valid; when we read the object after writing to it, we get the value 1, as expected.

We can ask FDR to verify the following refinements:

$\textrm{InternalConsistency} \mathrel{\sqsubseteq_{T}} \textrm{UnrepeatableRead}$

$\textrm{InternalConsistency} \mathrel{\sqsubseteq_{T}} \textrm{RepeatableRead}$

## Refinement

Model checking in CSP is all about refinement. $$Spec \mathrel{\sqsubseteq} Impl$$ says that $$Impl$$ refines $$Spec$$: all of the things that $$Impl$$ tries to do are “allowed” by $$Spec$$.

CSP also provides several semantic models, which define what kinds of behavior your $$Spec$$ process can allow or deny. Each semantic model has its own (incredibly mathematical) definition of what “refinement” means. The simplest semantic model is traces refinement ($$\sqsubseteq_{T}$$), which is normally used to verify a safety property — that is, that something bad is not allowed to occur. Later on we’ll learn about more complex semantic models that also let us check liveness properties — that is, that something good must occur.

I’ve wrapped up FDR in a Docker image in my csp-models repository; this repository also contains a machine-readable version of the internal consistency specification, and of the two refinements we want to check. If we run this example through FDR, we should get:

\$ ./refines concurrency-models/read-atomic.csp
Result: Failed
Counterexample (Trace Counterexample)
Specification Debug:
Trace: <write.1.1>
Implementation Debug:
Trace: <write.1.1>

Result: Passed


(Note that I haven’t included the entire output, just the important bits.)

We get the results that we expect: $$\textrm{UnrepeatableRead}$$ is invalid, and $$\textrm{RepeatableRead}$$ is valid. Even better, for the invalid transaction, FDR gives us a counterexample telling us what went wrong! Digging into the counterexample, we see that after the sequence of events $$\langle \texttt{write}{.}1{.}1 \rangle$$, our specification allows the following events:

• we can $$\texttt{write}$$ whatever we want;

• we can $$\texttt{read}$$ whatever we want from object 2 (since it’s still undefined);

• but if we want to $$\texttt{read}$$ from object 1, the only value we’re able to see is 1.

However, at this point our implementation wants to $$\textrm{read}$$ the value 2 from object 1. Since that’s not allowed by our specification, the refinement fails, and we have proof that $$\textrm{UnrepeatableRead}$$ does not satisfy $$\textrm{InternalConsistency}$$.

#### Next steps

That wraps it up for internal consistency! In the next couple of posts, we’ll take a look at external consistency, and then we’ll look at the reference Read Atomic implementation from the paper, and try to use FDR to show that the implementation really does implement Read Atomic.